Reference
How players are compared
Every comparison in this notebook is a preregistered match between two named
players on identical seeds, scored by effective win rate with a 95% interval.
The rules of the match, the players and the decision rule are frozen in
experiments/<uuid>.toml before a game is played.
Paired seeds
A match of n games uses n / 2 seeds. Each seed is played twice, once
with each player as Red, so that both sides see the same random setups and
the same tie-breaking randomness. The per-seed contrast (the candidate's
score as Red plus its score as Blue, minus one) is the unit of analysis for
paired experiments, and the reported interval is over seeds. Red moves
first; the Ataraxos self-play returns for Red and Blue were roughly equal,
which supports the human belief that move priority is not strategically
significant, but pairing removes the question from the comparison.
Effective win rate
A win counts one, a draw a half, a loss zero, following the scoring of the Ataraxos evaluation series. The record reports wins, draws and losses separately beside the effective rate, and the endings histogram (flag captured, no legal moves, both immobile, battleless-move limit, length limit, resignation) so a policy that wins by stalling is visible.
The 95% interval on a proportion is the Wilson score interval. For a paired contrast it is the normal-approximation interval over seeds. Neither is a significance test; the decision rule of each registration says what the interval has to exclude.
Rules presets
training is the paper's training configuration: a draw after 100
consecutive battleless moves and after 4,000 moves. competitive uses the
200-move rule of the Strategus platform. Both enforce the two-square rule
and the continuous-chase rule. A registration names its preset; results
under different presets are not pooled.
Elo and its caveats
Round-robin tournaments report Elo fitted from pairwise scores, anchored at
1000 for random. The Ataraxos paper warns that Elo in Stratego is useful
but flawed because randomisation compresses margins: two players can be
closer to each other than their margins against a common third player
predict. Elo here orders the ladder; promotion decisions use the paired
contrast against the previous rung, never an Elo difference.
Information boundary
A policy receives its own Observation and the legal moves computed by the
arbiter. The observation carries the public anti-chase trackers so that a
searcher can rebuild rule-exact states, but never a hidden type, the
arbiter's Zobrist key or the random seed. Search-based players sample
hidden pieces from the observation alone. The training environment does
expose hidden types, as belief targets, to the learner only.
What counts as support
A candidate is promoted when the preregistered interval on the paired contrast against the previous rung excludes zero in its favour on the full game count, and any cost constraint in the registration holds (time per move, memory). A run stopped early, an interval overlapping zero, or a result on a smoke preset is recorded and not promoted.